HAMMER VFS - Limit recursion for long directory chains, update mtime/ctime
[dragonfly.git] / sys / vfs / hammer / hammer_inode.c
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
35  */
36
37 #include "hammer.h"
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
41
42 static int      hammer_unload_inode(struct hammer_inode *ip);
43 static void     hammer_free_inode(hammer_inode_t ip);
44 static void     hammer_flush_inode_core(hammer_inode_t ip,
45                                         hammer_flush_group_t flg, int flags);
46 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
47 #if 0
48 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
49 #endif
50 static int      hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
51                                         hammer_flush_group_t flg);
52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
53                                         int depth, hammer_flush_group_t flg);
54 static void     hammer_inode_wakereclaims(hammer_inode_t ip, int dowake);
55
56 #ifdef DEBUG_TRUNCATE
57 extern struct hammer_inode *HammerTruncIp;
58 #endif
59
60 /*
61  * RB-Tree support for inode structures
62  */
63 int
64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
65 {
66         if (ip1->obj_localization < ip2->obj_localization)
67                 return(-1);
68         if (ip1->obj_localization > ip2->obj_localization)
69                 return(1);
70         if (ip1->obj_id < ip2->obj_id)
71                 return(-1);
72         if (ip1->obj_id > ip2->obj_id)
73                 return(1);
74         if (ip1->obj_asof < ip2->obj_asof)
75                 return(-1);
76         if (ip1->obj_asof > ip2->obj_asof)
77                 return(1);
78         return(0);
79 }
80
81 /*
82  * RB-Tree support for inode structures / special LOOKUP_INFO
83  */
84 static int
85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
86 {
87         if (info->obj_localization < ip->obj_localization)
88                 return(-1);
89         if (info->obj_localization > ip->obj_localization)
90                 return(1);
91         if (info->obj_id < ip->obj_id)
92                 return(-1);
93         if (info->obj_id > ip->obj_id)
94                 return(1);
95         if (info->obj_asof < ip->obj_asof)
96                 return(-1);
97         if (info->obj_asof > ip->obj_asof)
98                 return(1);
99         return(0);
100 }
101
102 /*
103  * Used by hammer_scan_inode_snapshots() to locate all of an object's
104  * snapshots.  Note that the asof field is not tested, which we can get
105  * away with because it is the lowest-priority field.
106  */
107 static int
108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
109 {
110         hammer_inode_info_t info = data;
111
112         if (ip->obj_localization > info->obj_localization)
113                 return(1);
114         if (ip->obj_localization < info->obj_localization)
115                 return(-1);
116         if (ip->obj_id > info->obj_id)
117                 return(1);
118         if (ip->obj_id < info->obj_id)
119                 return(-1);
120         return(0);
121 }
122
123 /*
124  * Used by hammer_unload_pseudofs() to locate all inodes associated with
125  * a particular PFS.
126  */
127 static int
128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
129 {
130         u_int32_t localization = *(u_int32_t *)data;
131         if (ip->obj_localization > localization)
132                 return(1);
133         if (ip->obj_localization < localization)
134                 return(-1);
135         return(0);
136 }
137
138 /*
139  * RB-Tree support for pseudofs structures
140  */
141 static int
142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
143 {
144         if (p1->localization < p2->localization)
145                 return(-1);
146         if (p1->localization > p2->localization)
147                 return(1);
148         return(0);
149 }
150
151
152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
154                 hammer_inode_info_cmp, hammer_inode_info_t);
155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
156              hammer_pfs_rb_compare, u_int32_t, localization);
157
158 /*
159  * The kernel is not actively referencing this vnode but is still holding
160  * it cached.
161  *
162  * This is called from the frontend.
163  */
164 int
165 hammer_vop_inactive(struct vop_inactive_args *ap)
166 {
167         struct hammer_inode *ip = VTOI(ap->a_vp);
168
169         /*
170          * Degenerate case
171          */
172         if (ip == NULL) {
173                 vrecycle(ap->a_vp);
174                 return(0);
175         }
176
177         /*
178          * If the inode no longer has visibility in the filesystem try to
179          * recycle it immediately, even if the inode is dirty.  Recycling
180          * it quickly allows the system to reclaim buffer cache and VM
181          * resources which can matter a lot in a heavily loaded system.
182          *
183          * This can deadlock in vfsync() if we aren't careful.
184          * 
185          * Do not queue the inode to the flusher if we still have visibility,
186          * otherwise namespace calls such as chmod will unnecessarily generate
187          * multiple inode updates.
188          */
189         hammer_inode_unloadable_check(ip, 0);
190         if (ip->ino_data.nlinks == 0) {
191                 if (ip->flags & HAMMER_INODE_MODMASK)
192                         hammer_flush_inode(ip, 0);
193                 vrecycle(ap->a_vp);
194         }
195         return(0);
196 }
197
198 /*
199  * Release the vnode association.  This is typically (but not always)
200  * the last reference on the inode.
201  *
202  * Once the association is lost we are on our own with regards to
203  * flushing the inode.
204  */
205 int
206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
207 {
208         struct hammer_inode *ip;
209         hammer_mount_t hmp;
210         struct vnode *vp;
211
212         vp = ap->a_vp;
213
214         if ((ip = vp->v_data) != NULL) {
215                 hmp = ip->hmp;
216                 vp->v_data = NULL;
217                 ip->vp = NULL;
218
219                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
220                         ++hammer_count_reclaiming;
221                         ++hmp->inode_reclaims;
222                         ip->flags |= HAMMER_INODE_RECLAIM;
223                 }
224                 hammer_rel_inode(ip, 1);
225         }
226         return(0);
227 }
228
229 /*
230  * Return a locked vnode for the specified inode.  The inode must be
231  * referenced but NOT LOCKED on entry and will remain referenced on
232  * return.
233  *
234  * Called from the frontend.
235  */
236 int
237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
238 {
239         hammer_mount_t hmp;
240         struct vnode *vp;
241         int error = 0;
242         u_int8_t obj_type;
243
244         hmp = ip->hmp;
245
246         for (;;) {
247                 if ((vp = ip->vp) == NULL) {
248                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
249                         if (error)
250                                 break;
251                         hammer_lock_ex(&ip->lock);
252                         if (ip->vp != NULL) {
253                                 hammer_unlock(&ip->lock);
254                                 vp->v_type = VBAD;
255                                 vx_put(vp);
256                                 continue;
257                         }
258                         hammer_ref(&ip->lock);
259                         vp = *vpp;
260                         ip->vp = vp;
261
262                         obj_type = ip->ino_data.obj_type;
263                         vp->v_type = hammer_get_vnode_type(obj_type);
264
265                         hammer_inode_wakereclaims(ip, 0);
266
267                         switch(ip->ino_data.obj_type) {
268                         case HAMMER_OBJTYPE_CDEV:
269                         case HAMMER_OBJTYPE_BDEV:
270                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
271                                 addaliasu(vp, ip->ino_data.rmajor,
272                                           ip->ino_data.rminor);
273                                 break;
274                         case HAMMER_OBJTYPE_FIFO:
275                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
276                                 break;
277                         default:
278                                 break;
279                         }
280
281                         /*
282                          * Only mark as the root vnode if the ip is not
283                          * historical, otherwise the VFS cache will get
284                          * confused.  The other half of the special handling
285                          * is in hammer_vop_nlookupdotdot().
286                          *
287                          * Pseudo-filesystem roots can be accessed via
288                          * non-root filesystem paths and setting VROOT may
289                          * confuse the namecache.  Set VPFSROOT instead.
290                          */
291                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
292                             ip->obj_asof == hmp->asof) {
293                                 if (ip->obj_localization == 0)
294                                         vp->v_flag |= VROOT;
295                                 else
296                                         vp->v_flag |= VPFSROOT;
297                         }
298
299                         vp->v_data = (void *)ip;
300                         /* vnode locked by getnewvnode() */
301                         /* make related vnode dirty if inode dirty? */
302                         hammer_unlock(&ip->lock);
303                         if (vp->v_type == VREG)
304                                 vinitvmio(vp, ip->ino_data.size);
305                         break;
306                 }
307
308                 /*
309                  * loop if the vget fails (aka races), or if the vp
310                  * no longer matches ip->vp.
311                  */
312                 if (vget(vp, LK_EXCLUSIVE) == 0) {
313                         if (vp == ip->vp)
314                                 break;
315                         vput(vp);
316                 }
317         }
318         *vpp = vp;
319         return(error);
320 }
321
322 /*
323  * Locate all copies of the inode for obj_id compatible with the specified
324  * asof, reference, and issue the related call-back.  This routine is used
325  * for direct-io invalidation and does not create any new inodes.
326  */
327 void
328 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
329                             int (*callback)(hammer_inode_t ip, void *data),
330                             void *data)
331 {
332         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
333                                    hammer_inode_info_cmp_all_history,
334                                    callback, iinfo);
335 }
336
337 /*
338  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
339  * do not attach or detach the related vnode (use hammer_get_vnode() for
340  * that).
341  *
342  * The flags argument is only applied for newly created inodes, and only
343  * certain flags are inherited.
344  *
345  * Called from the frontend.
346  */
347 struct hammer_inode *
348 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
349                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
350                  int flags, int *errorp)
351 {
352         hammer_mount_t hmp = trans->hmp;
353         struct hammer_inode_info iinfo;
354         struct hammer_cursor cursor;
355         struct hammer_inode *ip;
356
357
358         /*
359          * Determine if we already have an inode cached.  If we do then
360          * we are golden.
361          *
362          * If we find an inode with no vnode we have to mark the
363          * transaction such that hammer_inode_waitreclaims() is
364          * called later on to avoid building up an infinite number
365          * of inodes.  Otherwise we can continue to * add new inodes
366          * faster then they can be disposed of, even with the tsleep
367          * delay.
368          *
369          * If we find a dummy inode we return a failure so dounlink
370          * (which does another lookup) doesn't try to mess with the
371          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
372          * to ref dummy inodes.
373          */
374         iinfo.obj_id = obj_id;
375         iinfo.obj_asof = asof;
376         iinfo.obj_localization = localization;
377 loop:
378         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
379         if (ip) {
380                 if (ip->flags & HAMMER_INODE_DUMMY) {
381                         *errorp = ENOENT;
382                         return(NULL);
383                 }
384                 hammer_ref(&ip->lock);
385                 *errorp = 0;
386                 return(ip);
387         }
388
389         /*
390          * Allocate a new inode structure and deal with races later.
391          */
392         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
393         ++hammer_count_inodes;
394         ++hmp->count_inodes;
395         ip->obj_id = obj_id;
396         ip->obj_asof = iinfo.obj_asof;
397         ip->obj_localization = localization;
398         ip->hmp = hmp;
399         ip->flags = flags & HAMMER_INODE_RO;
400         ip->cache[0].ip = ip;
401         ip->cache[1].ip = ip;
402         if (hmp->ronly)
403                 ip->flags |= HAMMER_INODE_RO;
404         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
405                 0x7FFFFFFFFFFFFFFFLL;
406         RB_INIT(&ip->rec_tree);
407         TAILQ_INIT(&ip->target_list);
408         hammer_ref(&ip->lock);
409
410         /*
411          * Locate the on-disk inode.  If this is a PFS root we always
412          * access the current version of the root inode and (if it is not
413          * a master) always access information under it with a snapshot
414          * TID.
415          */
416 retry:
417         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
418         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
419         cursor.key_beg.obj_id = ip->obj_id;
420         cursor.key_beg.key = 0;
421         cursor.key_beg.create_tid = 0;
422         cursor.key_beg.delete_tid = 0;
423         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
424         cursor.key_beg.obj_type = 0;
425
426         cursor.asof = iinfo.obj_asof;
427         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
428                        HAMMER_CURSOR_ASOF;
429
430         *errorp = hammer_btree_lookup(&cursor);
431         if (*errorp == EDEADLK) {
432                 hammer_done_cursor(&cursor);
433                 goto retry;
434         }
435
436         /*
437          * On success the B-Tree lookup will hold the appropriate
438          * buffer cache buffers and provide a pointer to the requested
439          * information.  Copy the information to the in-memory inode
440          * and cache the B-Tree node to improve future operations.
441          */
442         if (*errorp == 0) {
443                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
444                 ip->ino_data = cursor.data->inode;
445
446                 /*
447                  * cache[0] tries to cache the location of the object inode.
448                  * The assumption is that it is near the directory inode.
449                  *
450                  * cache[1] tries to cache the location of the object data.
451                  * The assumption is that it is near the directory data.
452                  */
453                 hammer_cache_node(&ip->cache[0], cursor.node);
454                 if (dip && dip->cache[1].node)
455                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
456
457                 /*
458                  * The file should not contain any data past the file size
459                  * stored in the inode.  Setting save_trunc_off to the
460                  * file size instead of max reduces B-Tree lookup overheads
461                  * on append by allowing the flusher to avoid checking for
462                  * record overwrites.
463                  */
464                 ip->save_trunc_off = ip->ino_data.size;
465
466                 /*
467                  * Locate and assign the pseudofs management structure to
468                  * the inode.
469                  */
470                 if (dip && dip->obj_localization == ip->obj_localization) {
471                         ip->pfsm = dip->pfsm;
472                         hammer_ref(&ip->pfsm->lock);
473                 } else {
474                         ip->pfsm = hammer_load_pseudofs(trans,
475                                                         ip->obj_localization,
476                                                         errorp);
477                         *errorp = 0;    /* ignore ENOENT */
478                 }
479         }
480
481         /*
482          * The inode is placed on the red-black tree and will be synced to
483          * the media when flushed or by the filesystem sync.  If this races
484          * another instantiation/lookup the insertion will fail.
485          */
486         if (*errorp == 0) {
487                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
488                         hammer_free_inode(ip);
489                         hammer_done_cursor(&cursor);
490                         goto loop;
491                 }
492                 ip->flags |= HAMMER_INODE_ONDISK;
493         } else {
494                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
495                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
496                         --hmp->rsv_inodes;
497                 }
498
499                 hammer_free_inode(ip);
500                 ip = NULL;
501         }
502         hammer_done_cursor(&cursor);
503         trans->flags |= HAMMER_TRANSF_NEWINODE;
504         return (ip);
505 }
506
507 /*
508  * Get a dummy inode to placemark a broken directory entry.
509  */
510 struct hammer_inode *
511 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
512                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
513                  int flags, int *errorp)
514 {
515         hammer_mount_t hmp = trans->hmp;
516         struct hammer_inode_info iinfo;
517         struct hammer_inode *ip;
518
519         /*
520          * Determine if we already have an inode cached.  If we do then
521          * we are golden.
522          *
523          * If we find an inode with no vnode we have to mark the
524          * transaction such that hammer_inode_waitreclaims() is
525          * called later on to avoid building up an infinite number
526          * of inodes.  Otherwise we can continue to * add new inodes
527          * faster then they can be disposed of, even with the tsleep
528          * delay.
529          *
530          * If we find a non-fake inode we return an error.  Only fake
531          * inodes can be returned by this routine.
532          */
533         iinfo.obj_id = obj_id;
534         iinfo.obj_asof = asof;
535         iinfo.obj_localization = localization;
536 loop:
537         *errorp = 0;
538         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
539         if (ip) {
540                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
541                         *errorp = ENOENT;
542                         return(NULL);
543                 }
544                 hammer_ref(&ip->lock);
545                 return(ip);
546         }
547
548         /*
549          * Allocate a new inode structure and deal with races later.
550          */
551         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
552         ++hammer_count_inodes;
553         ++hmp->count_inodes;
554         ip->obj_id = obj_id;
555         ip->obj_asof = iinfo.obj_asof;
556         ip->obj_localization = localization;
557         ip->hmp = hmp;
558         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
559         ip->cache[0].ip = ip;
560         ip->cache[1].ip = ip;
561         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
562                 0x7FFFFFFFFFFFFFFFLL;
563         RB_INIT(&ip->rec_tree);
564         TAILQ_INIT(&ip->target_list);
565         hammer_ref(&ip->lock);
566
567         /*
568          * Populate the dummy inode.  Leave everything zero'd out.
569          *
570          * (ip->ino_leaf and ip->ino_data)
571          *
572          * Make the dummy inode a FIFO object which most copy programs
573          * will properly ignore.
574          */
575         ip->save_trunc_off = ip->ino_data.size;
576         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
577
578         /*
579          * Locate and assign the pseudofs management structure to
580          * the inode.
581          */
582         if (dip && dip->obj_localization == ip->obj_localization) {
583                 ip->pfsm = dip->pfsm;
584                 hammer_ref(&ip->pfsm->lock);
585         } else {
586                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
587                                                 errorp);
588                 *errorp = 0;    /* ignore ENOENT */
589         }
590
591         /*
592          * The inode is placed on the red-black tree and will be synced to
593          * the media when flushed or by the filesystem sync.  If this races
594          * another instantiation/lookup the insertion will fail.
595          *
596          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
597          */
598         if (*errorp == 0) {
599                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
600                         hammer_free_inode(ip);
601                         goto loop;
602                 }
603         } else {
604                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
605                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
606                         --hmp->rsv_inodes;
607                 }
608                 hammer_free_inode(ip);
609                 ip = NULL;
610         }
611         trans->flags |= HAMMER_TRANSF_NEWINODE;
612         return (ip);
613 }
614
615 /*
616  * Create a new filesystem object, returning the inode in *ipp.  The
617  * returned inode will be referenced.  The inode is created in-memory.
618  *
619  * If pfsm is non-NULL the caller wishes to create the root inode for
620  * a master PFS.
621  */
622 int
623 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
624                     struct ucred *cred, hammer_inode_t dip,
625                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
626 {
627         hammer_mount_t hmp;
628         hammer_inode_t ip;
629         uid_t xuid;
630         int error;
631
632         hmp = trans->hmp;
633
634         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
635         ++hammer_count_inodes;
636         ++hmp->count_inodes;
637         trans->flags |= HAMMER_TRANSF_NEWINODE;
638
639         if (pfsm) {
640                 KKASSERT(pfsm->localization != 0);
641                 ip->obj_id = HAMMER_OBJID_ROOT;
642                 ip->obj_localization = pfsm->localization;
643         } else {
644                 KKASSERT(dip != NULL);
645                 ip->obj_id = hammer_alloc_objid(hmp, dip);
646                 ip->obj_localization = dip->obj_localization;
647         }
648
649         KKASSERT(ip->obj_id != 0);
650         ip->obj_asof = hmp->asof;
651         ip->hmp = hmp;
652         ip->flush_state = HAMMER_FST_IDLE;
653         ip->flags = HAMMER_INODE_DDIRTY |
654                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
655         ip->cache[0].ip = ip;
656         ip->cache[1].ip = ip;
657
658         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
659         /* ip->save_trunc_off = 0; (already zero) */
660         RB_INIT(&ip->rec_tree);
661         TAILQ_INIT(&ip->target_list);
662
663         ip->ino_data.atime = trans->time;
664         ip->ino_data.mtime = trans->time;
665         ip->ino_data.size = 0;
666         ip->ino_data.nlinks = 0;
667
668         /*
669          * A nohistory designator on the parent directory is inherited by
670          * the child.  We will do this even for pseudo-fs creation... the
671          * sysad can turn it off.
672          */
673         if (dip) {
674                 ip->ino_data.uflags = dip->ino_data.uflags &
675                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
676         }
677
678         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
679         ip->ino_leaf.base.localization = ip->obj_localization +
680                                          HAMMER_LOCALIZE_INODE;
681         ip->ino_leaf.base.obj_id = ip->obj_id;
682         ip->ino_leaf.base.key = 0;
683         ip->ino_leaf.base.create_tid = 0;
684         ip->ino_leaf.base.delete_tid = 0;
685         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
686         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
687
688         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
689         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
690         ip->ino_data.mode = vap->va_mode;
691         ip->ino_data.ctime = trans->time;
692
693         /*
694          * If we are running version 2 or greater we use dirhash algorithm #1
695          * which is semi-sorted.  Algorithm #0 was just a pure crc.
696          */
697         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
698                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
699                         ip->ino_data.cap_flags |= HAMMER_INODE_CAP_DIRHASH_ALG1;
700                 }
701         }
702
703         /*
704          * Setup the ".." pointer.  This only needs to be done for directories
705          * but we do it for all objects as a recovery aid.
706          */
707         if (dip)
708                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
709 #if 0
710         /*
711          * The parent_obj_localization field only applies to pseudo-fs roots.
712          * XXX this is no longer applicable, PFSs are no longer directly
713          * tied into the parent's directory structure.
714          */
715         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
716             ip->obj_id == HAMMER_OBJID_ROOT) {
717                 ip->ino_data.ext.obj.parent_obj_localization = 
718                                                 dip->obj_localization;
719         }
720 #endif
721
722         switch(ip->ino_leaf.base.obj_type) {
723         case HAMMER_OBJTYPE_CDEV:
724         case HAMMER_OBJTYPE_BDEV:
725                 ip->ino_data.rmajor = vap->va_rmajor;
726                 ip->ino_data.rminor = vap->va_rminor;
727                 break;
728         default:
729                 break;
730         }
731
732         /*
733          * Calculate default uid/gid and overwrite with information from
734          * the vap.
735          */
736         if (dip) {
737                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
738                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
739                                              xuid, cred, &vap->va_mode);
740         } else {
741                 xuid = 0;
742         }
743         ip->ino_data.mode = vap->va_mode;
744
745         if (vap->va_vaflags & VA_UID_UUID_VALID)
746                 ip->ino_data.uid = vap->va_uid_uuid;
747         else if (vap->va_uid != (uid_t)VNOVAL)
748                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
749         else
750                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
751
752         if (vap->va_vaflags & VA_GID_UUID_VALID)
753                 ip->ino_data.gid = vap->va_gid_uuid;
754         else if (vap->va_gid != (gid_t)VNOVAL)
755                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
756         else if (dip)
757                 ip->ino_data.gid = dip->ino_data.gid;
758
759         hammer_ref(&ip->lock);
760
761         if (pfsm) {
762                 ip->pfsm = pfsm;
763                 hammer_ref(&pfsm->lock);
764                 error = 0;
765         } else if (dip->obj_localization == ip->obj_localization) {
766                 ip->pfsm = dip->pfsm;
767                 hammer_ref(&ip->pfsm->lock);
768                 error = 0;
769         } else {
770                 ip->pfsm = hammer_load_pseudofs(trans,
771                                                 ip->obj_localization,
772                                                 &error);
773                 error = 0;      /* ignore ENOENT */
774         }
775
776         if (error) {
777                 hammer_free_inode(ip);
778                 ip = NULL;
779         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
780                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
781                 /* not reached */
782                 hammer_free_inode(ip);
783         }
784         *ipp = ip;
785         return(error);
786 }
787
788 /*
789  * Final cleanup / freeing of an inode structure
790  */
791 static void
792 hammer_free_inode(hammer_inode_t ip)
793 {
794         struct hammer_mount *hmp;
795
796         hmp = ip->hmp;
797         KKASSERT(ip->lock.refs == 1);
798         hammer_uncache_node(&ip->cache[0]);
799         hammer_uncache_node(&ip->cache[1]);
800         hammer_inode_wakereclaims(ip, 1);
801         if (ip->objid_cache)
802                 hammer_clear_objid(ip);
803         --hammer_count_inodes;
804         --hmp->count_inodes;
805         if (ip->pfsm) {
806                 hammer_rel_pseudofs(hmp, ip->pfsm);
807                 ip->pfsm = NULL;
808         }
809         kfree(ip, hmp->m_inodes);
810         ip = NULL;
811 }
812
813 /*
814  * Retrieve pseudo-fs data.  NULL will never be returned.
815  *
816  * If an error occurs *errorp will be set and a default template is returned,
817  * otherwise *errorp is set to 0.  Typically when an error occurs it will
818  * be ENOENT.
819  */
820 hammer_pseudofs_inmem_t
821 hammer_load_pseudofs(hammer_transaction_t trans,
822                      u_int32_t localization, int *errorp)
823 {
824         hammer_mount_t hmp = trans->hmp;
825         hammer_inode_t ip;
826         hammer_pseudofs_inmem_t pfsm;
827         struct hammer_cursor cursor;
828         int bytes;
829
830 retry:
831         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
832         if (pfsm) {
833                 hammer_ref(&pfsm->lock);
834                 *errorp = 0;
835                 return(pfsm);
836         }
837
838         /*
839          * PFS records are stored in the root inode (not the PFS root inode,
840          * but the real root).  Avoid an infinite recursion if loading
841          * the PFS for the real root.
842          */
843         if (localization) {
844                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
845                                       HAMMER_MAX_TID,
846                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
847         } else {
848                 ip = NULL;
849         }
850
851         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
852         pfsm->localization = localization;
853         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
854         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
855
856         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
857         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
858                                       HAMMER_LOCALIZE_MISC;
859         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
860         cursor.key_beg.create_tid = 0;
861         cursor.key_beg.delete_tid = 0;
862         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
863         cursor.key_beg.obj_type = 0;
864         cursor.key_beg.key = localization;
865         cursor.asof = HAMMER_MAX_TID;
866         cursor.flags |= HAMMER_CURSOR_ASOF;
867
868         if (ip)
869                 *errorp = hammer_ip_lookup(&cursor);
870         else
871                 *errorp = hammer_btree_lookup(&cursor);
872         if (*errorp == 0) {
873                 *errorp = hammer_ip_resolve_data(&cursor);
874                 if (*errorp == 0) {
875                         if (cursor.data->pfsd.mirror_flags &
876                             HAMMER_PFSD_DELETED) {
877                                 *errorp = ENOENT;
878                         } else {
879                                 bytes = cursor.leaf->data_len;
880                                 if (bytes > sizeof(pfsm->pfsd))
881                                         bytes = sizeof(pfsm->pfsd);
882                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
883                         }
884                 }
885         }
886         hammer_done_cursor(&cursor);
887
888         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
889         hammer_ref(&pfsm->lock);
890         if (ip)
891                 hammer_rel_inode(ip, 0);
892         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
893                 kfree(pfsm, hmp->m_misc);
894                 goto retry;
895         }
896         return(pfsm);
897 }
898
899 /*
900  * Store pseudo-fs data.  The backend will automatically delete any prior
901  * on-disk pseudo-fs data but we have to delete in-memory versions.
902  */
903 int
904 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
905 {
906         struct hammer_cursor cursor;
907         hammer_record_t record;
908         hammer_inode_t ip;
909         int error;
910
911         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
912                               HAMMER_DEF_LOCALIZATION, 0, &error);
913 retry:
914         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
915         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
916         cursor.key_beg.localization = ip->obj_localization +
917                                       HAMMER_LOCALIZE_MISC;
918         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
919         cursor.key_beg.create_tid = 0;
920         cursor.key_beg.delete_tid = 0;
921         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
922         cursor.key_beg.obj_type = 0;
923         cursor.key_beg.key = pfsm->localization;
924         cursor.asof = HAMMER_MAX_TID;
925         cursor.flags |= HAMMER_CURSOR_ASOF;
926
927         /*
928          * Replace any in-memory version of the record.
929          */
930         error = hammer_ip_lookup(&cursor);
931         if (error == 0 && hammer_cursor_inmem(&cursor)) {
932                 record = cursor.iprec;
933                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
934                         KKASSERT(cursor.deadlk_rec == NULL);
935                         hammer_ref(&record->lock);
936                         cursor.deadlk_rec = record;
937                         error = EDEADLK;
938                 } else {
939                         record->flags |= HAMMER_RECF_DELETED_FE;
940                         error = 0;
941                 }
942         }
943
944         /*
945          * Allocate replacement general record.  The backend flush will
946          * delete any on-disk version of the record.
947          */
948         if (error == 0 || error == ENOENT) {
949                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
950                 record->type = HAMMER_MEM_RECORD_GENERAL;
951
952                 record->leaf.base.localization = ip->obj_localization +
953                                                  HAMMER_LOCALIZE_MISC;
954                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
955                 record->leaf.base.key = pfsm->localization;
956                 record->leaf.data_len = sizeof(pfsm->pfsd);
957                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
958                 error = hammer_ip_add_record(trans, record);
959         }
960         hammer_done_cursor(&cursor);
961         if (error == EDEADLK)
962                 goto retry;
963         hammer_rel_inode(ip, 0);
964         return(error);
965 }
966
967 /*
968  * Create a root directory for a PFS if one does not alredy exist.
969  *
970  * The PFS root stands alone so we must also bump the nlinks count
971  * to prevent it from being destroyed on release.
972  */
973 int
974 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
975                        hammer_pseudofs_inmem_t pfsm)
976 {
977         hammer_inode_t ip;
978         struct vattr vap;
979         int error;
980
981         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
982                               pfsm->localization, 0, &error);
983         if (ip == NULL) {
984                 vattr_null(&vap);
985                 vap.va_mode = 0755;
986                 vap.va_type = VDIR;
987                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
988                 if (error == 0) {
989                         ++ip->ino_data.nlinks;
990                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
991                 }
992         }
993         if (ip)
994                 hammer_rel_inode(ip, 0);
995         return(error);
996 }
997
998 /*
999  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1000  * if we are unable to disassociate all the inodes.
1001  */
1002 static
1003 int
1004 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1005 {
1006         int res;
1007
1008         hammer_ref(&ip->lock);
1009         if (ip->lock.refs == 2 && ip->vp)
1010                 vclean_unlocked(ip->vp);
1011         if (ip->lock.refs == 1 && ip->vp == NULL)
1012                 res = 0;
1013         else
1014                 res = -1;       /* stop, someone is using the inode */
1015         hammer_rel_inode(ip, 0);
1016         return(res);
1017 }
1018
1019 int
1020 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1021 {
1022         int res;
1023         int try;
1024
1025         for (try = res = 0; try < 4; ++try) {
1026                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1027                                            hammer_inode_pfs_cmp,
1028                                            hammer_unload_pseudofs_callback,
1029                                            &localization);
1030                 if (res == 0 && try > 1)
1031                         break;
1032                 hammer_flusher_sync(trans->hmp);
1033         }
1034         if (res != 0)
1035                 res = ENOTEMPTY;
1036         return(res);
1037 }
1038
1039
1040 /*
1041  * Release a reference on a PFS
1042  */
1043 void
1044 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1045 {
1046         hammer_unref(&pfsm->lock);
1047         if (pfsm->lock.refs == 0) {
1048                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1049                 kfree(pfsm, hmp->m_misc);
1050         }
1051 }
1052
1053 /*
1054  * Called by hammer_sync_inode().
1055  */
1056 static int
1057 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1058 {
1059         hammer_transaction_t trans = cursor->trans;
1060         hammer_record_t record;
1061         int error;
1062         int redirty;
1063
1064 retry:
1065         error = 0;
1066
1067         /*
1068          * If the inode has a presence on-disk then locate it and mark
1069          * it deleted, setting DELONDISK.
1070          *
1071          * The record may or may not be physically deleted, depending on
1072          * the retention policy.
1073          */
1074         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1075             HAMMER_INODE_ONDISK) {
1076                 hammer_normalize_cursor(cursor);
1077                 cursor->key_beg.localization = ip->obj_localization + 
1078                                                HAMMER_LOCALIZE_INODE;
1079                 cursor->key_beg.obj_id = ip->obj_id;
1080                 cursor->key_beg.key = 0;
1081                 cursor->key_beg.create_tid = 0;
1082                 cursor->key_beg.delete_tid = 0;
1083                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1084                 cursor->key_beg.obj_type = 0;
1085                 cursor->asof = ip->obj_asof;
1086                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1087                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1088                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1089
1090                 error = hammer_btree_lookup(cursor);
1091                 if (hammer_debug_inode)
1092                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1093
1094                 if (error == 0) {
1095                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1096                         if (hammer_debug_inode)
1097                                 kprintf(" error %d\n", error);
1098                         if (error == 0) {
1099                                 ip->flags |= HAMMER_INODE_DELONDISK;
1100                         }
1101                         if (cursor->node)
1102                                 hammer_cache_node(&ip->cache[0], cursor->node);
1103                 }
1104                 if (error == EDEADLK) {
1105                         hammer_done_cursor(cursor);
1106                         error = hammer_init_cursor(trans, cursor,
1107                                                    &ip->cache[0], ip);
1108                         if (hammer_debug_inode)
1109                                 kprintf("IPDED %p %d\n", ip, error);
1110                         if (error == 0)
1111                                 goto retry;
1112                 }
1113         }
1114
1115         /*
1116          * Ok, write out the initial record or a new record (after deleting
1117          * the old one), unless the DELETED flag is set.  This routine will
1118          * clear DELONDISK if it writes out a record.
1119          *
1120          * Update our inode statistics if this is the first application of
1121          * the inode on-disk.
1122          */
1123         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1124                 /*
1125                  * Generate a record and write it to the media.  We clean-up
1126                  * the state before releasing so we do not have to set-up
1127                  * a flush_group.
1128                  */
1129                 record = hammer_alloc_mem_record(ip, 0);
1130                 record->type = HAMMER_MEM_RECORD_INODE;
1131                 record->flush_state = HAMMER_FST_FLUSH;
1132                 record->leaf = ip->sync_ino_leaf;
1133                 record->leaf.base.create_tid = trans->tid;
1134                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1135                 record->leaf.create_ts = trans->time32;
1136                 record->data = (void *)&ip->sync_ino_data;
1137                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1138
1139                 /*
1140                  * If this flag is set we cannot sync the new file size
1141                  * because we haven't finished related truncations.  The
1142                  * inode will be flushed in another flush group to finish
1143                  * the job.
1144                  */
1145                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1146                     ip->sync_ino_data.size != ip->ino_data.size) {
1147                         redirty = 1;
1148                         ip->sync_ino_data.size = ip->ino_data.size;
1149                 } else {
1150                         redirty = 0;
1151                 }
1152
1153                 for (;;) {
1154                         error = hammer_ip_sync_record_cursor(cursor, record);
1155                         if (hammer_debug_inode)
1156                                 kprintf("GENREC %p rec %08x %d\n",      
1157                                         ip, record->flags, error);
1158                         if (error != EDEADLK)
1159                                 break;
1160                         hammer_done_cursor(cursor);
1161                         error = hammer_init_cursor(trans, cursor,
1162                                                    &ip->cache[0], ip);
1163                         if (hammer_debug_inode)
1164                                 kprintf("GENREC reinit %d\n", error);
1165                         if (error)
1166                                 break;
1167                 }
1168
1169                 /*
1170                  * Note:  The record was never on the inode's record tree
1171                  * so just wave our hands importantly and destroy it.
1172                  */
1173                 record->flags |= HAMMER_RECF_COMMITTED;
1174                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1175                 record->flush_state = HAMMER_FST_IDLE;
1176                 ++ip->rec_generation;
1177                 hammer_rel_mem_record(record);
1178
1179                 /*
1180                  * Finish up.
1181                  */
1182                 if (error == 0) {
1183                         if (hammer_debug_inode)
1184                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1185                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1186                                             HAMMER_INODE_ATIME |
1187                                             HAMMER_INODE_MTIME);
1188                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1189                         if (redirty)
1190                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1191
1192                         /*
1193                          * Root volume count of inodes
1194                          */
1195                         hammer_sync_lock_sh(trans);
1196                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1197                                 hammer_modify_volume_field(trans,
1198                                                            trans->rootvol,
1199                                                            vol0_stat_inodes);
1200                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1201                                 hammer_modify_volume_done(trans->rootvol);
1202                                 ip->flags |= HAMMER_INODE_ONDISK;
1203                                 if (hammer_debug_inode)
1204                                         kprintf("NOWONDISK %p\n", ip);
1205                         }
1206                         hammer_sync_unlock(trans);
1207                 }
1208         }
1209
1210         /*
1211          * If the inode has been destroyed, clean out any left-over flags
1212          * that may have been set by the frontend.
1213          */
1214         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1215                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1216                                     HAMMER_INODE_ATIME |
1217                                     HAMMER_INODE_MTIME);
1218         }
1219         return(error);
1220 }
1221
1222 /*
1223  * Update only the itimes fields.
1224  *
1225  * ATIME can be updated without generating any UNDO.  MTIME is updated
1226  * with UNDO so it is guaranteed to be synchronized properly in case of
1227  * a crash.
1228  *
1229  * Neither field is included in the B-Tree leaf element's CRC, which is how
1230  * we can get away with updating ATIME the way we do.
1231  */
1232 static int
1233 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1234 {
1235         hammer_transaction_t trans = cursor->trans;
1236         int error;
1237
1238 retry:
1239         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1240             HAMMER_INODE_ONDISK) {
1241                 return(0);
1242         }
1243
1244         hammer_normalize_cursor(cursor);
1245         cursor->key_beg.localization = ip->obj_localization + 
1246                                        HAMMER_LOCALIZE_INODE;
1247         cursor->key_beg.obj_id = ip->obj_id;
1248         cursor->key_beg.key = 0;
1249         cursor->key_beg.create_tid = 0;
1250         cursor->key_beg.delete_tid = 0;
1251         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1252         cursor->key_beg.obj_type = 0;
1253         cursor->asof = ip->obj_asof;
1254         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1255         cursor->flags |= HAMMER_CURSOR_ASOF;
1256         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1257         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1258         cursor->flags |= HAMMER_CURSOR_BACKEND;
1259
1260         error = hammer_btree_lookup(cursor);
1261         if (error == 0) {
1262                 hammer_cache_node(&ip->cache[0], cursor->node);
1263                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1264                         /*
1265                          * Updating MTIME requires an UNDO.  Just cover
1266                          * both atime and mtime.
1267                          */
1268                         hammer_sync_lock_sh(trans);
1269                         hammer_modify_buffer(trans, cursor->data_buffer,
1270                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1271                                      HAMMER_ITIMES_BYTES);
1272                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1273                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1274                         hammer_modify_buffer_done(cursor->data_buffer);
1275                         hammer_sync_unlock(trans);
1276                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1277                         /*
1278                          * Updating atime only can be done in-place with
1279                          * no UNDO.
1280                          */
1281                         hammer_sync_lock_sh(trans);
1282                         hammer_modify_buffer(trans, cursor->data_buffer,
1283                                              NULL, 0);
1284                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1285                         hammer_modify_buffer_done(cursor->data_buffer);
1286                         hammer_sync_unlock(trans);
1287                 }
1288                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1289         }
1290         if (error == EDEADLK) {
1291                 hammer_done_cursor(cursor);
1292                 error = hammer_init_cursor(trans, cursor,
1293                                            &ip->cache[0], ip);
1294                 if (error == 0)
1295                         goto retry;
1296         }
1297         return(error);
1298 }
1299
1300 /*
1301  * Release a reference on an inode, flush as requested.
1302  *
1303  * On the last reference we queue the inode to the flusher for its final
1304  * disposition.
1305  */
1306 void
1307 hammer_rel_inode(struct hammer_inode *ip, int flush)
1308 {
1309         /*hammer_mount_t hmp = ip->hmp;*/
1310
1311         /*
1312          * Handle disposition when dropping the last ref.
1313          */
1314         for (;;) {
1315                 if (ip->lock.refs == 1) {
1316                         /*
1317                          * Determine whether on-disk action is needed for
1318                          * the inode's final disposition.
1319                          */
1320                         KKASSERT(ip->vp == NULL);
1321                         hammer_inode_unloadable_check(ip, 0);
1322                         if (ip->flags & HAMMER_INODE_MODMASK) {
1323                                 hammer_flush_inode(ip, 0);
1324                         } else if (ip->lock.refs == 1) {
1325                                 hammer_unload_inode(ip);
1326                                 break;
1327                         }
1328                 } else {
1329                         if (flush)
1330                                 hammer_flush_inode(ip, 0);
1331
1332                         /*
1333                          * The inode still has multiple refs, try to drop
1334                          * one ref.
1335                          */
1336                         KKASSERT(ip->lock.refs >= 1);
1337                         if (ip->lock.refs > 1) {
1338                                 hammer_unref(&ip->lock);
1339                                 break;
1340                         }
1341                 }
1342         }
1343 }
1344
1345 /*
1346  * Unload and destroy the specified inode.  Must be called with one remaining
1347  * reference.  The reference is disposed of.
1348  *
1349  * The inode must be completely clean.
1350  */
1351 static int
1352 hammer_unload_inode(struct hammer_inode *ip)
1353 {
1354         hammer_mount_t hmp = ip->hmp;
1355
1356         KASSERT(ip->lock.refs == 1,
1357                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1358         KKASSERT(ip->vp == NULL);
1359         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1360         KKASSERT(ip->cursor_ip_refs == 0);
1361         KKASSERT(ip->lock.lockcount == 0);
1362         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1363
1364         KKASSERT(RB_EMPTY(&ip->rec_tree));
1365         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1366
1367         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1368
1369         hammer_free_inode(ip);
1370         return(0);
1371 }
1372
1373 /*
1374  * Called during unmounting if a critical error occured.  The in-memory
1375  * inode and all related structures are destroyed.
1376  *
1377  * If a critical error did not occur the unmount code calls the standard
1378  * release and asserts that the inode is gone.
1379  */
1380 int
1381 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1382 {
1383         hammer_record_t rec;
1384
1385         /*
1386          * Get rid of the inodes in-memory records, regardless of their
1387          * state, and clear the mod-mask.
1388          */
1389         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1390                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1391                 rec->target_ip = NULL;
1392                 if (rec->flush_state == HAMMER_FST_SETUP)
1393                         rec->flush_state = HAMMER_FST_IDLE;
1394         }
1395         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1396                 if (rec->flush_state == HAMMER_FST_FLUSH)
1397                         --rec->flush_group->refs;
1398                 else
1399                         hammer_ref(&rec->lock);
1400                 KKASSERT(rec->lock.refs == 1);
1401                 rec->flush_state = HAMMER_FST_IDLE;
1402                 rec->flush_group = NULL;
1403                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1404                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1405                 ++ip->rec_generation;
1406                 hammer_rel_mem_record(rec);
1407         }
1408         ip->flags &= ~HAMMER_INODE_MODMASK;
1409         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1410         KKASSERT(ip->vp == NULL);
1411
1412         /*
1413          * Remove the inode from any flush group, force it idle.  FLUSH
1414          * and SETUP states have an inode ref.
1415          */
1416         switch(ip->flush_state) {
1417         case HAMMER_FST_FLUSH:
1418                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1419                 --ip->flush_group->refs;
1420                 ip->flush_group = NULL;
1421                 /* fall through */
1422         case HAMMER_FST_SETUP:
1423                 hammer_unref(&ip->lock);
1424                 ip->flush_state = HAMMER_FST_IDLE;
1425                 /* fall through */
1426         case HAMMER_FST_IDLE:
1427                 break;
1428         }
1429
1430         /*
1431          * There shouldn't be any associated vnode.  The unload needs at
1432          * least one ref, if we do have a vp steal its ip ref.
1433          */
1434         if (ip->vp) {
1435                 kprintf("hammer_destroy_inode_callback: Unexpected "
1436                         "vnode association ip %p vp %p\n", ip, ip->vp);
1437                 ip->vp->v_data = NULL;
1438                 ip->vp = NULL;
1439         } else {
1440                 hammer_ref(&ip->lock);
1441         }
1442         hammer_unload_inode(ip);
1443         return(0);
1444 }
1445
1446 /*
1447  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1448  * the read-only flag for cached inodes.
1449  *
1450  * This routine is called from a RB_SCAN().
1451  */
1452 int
1453 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1454 {
1455         hammer_mount_t hmp = ip->hmp;
1456
1457         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1458                 ip->flags |= HAMMER_INODE_RO;
1459         else
1460                 ip->flags &= ~HAMMER_INODE_RO;
1461         return(0);
1462 }
1463
1464 /*
1465  * A transaction has modified an inode, requiring updates as specified by
1466  * the passed flags.
1467  *
1468  * HAMMER_INODE_DDIRTY: Inode data has been updated
1469  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1470  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1471  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1472  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1473  */
1474 void
1475 hammer_modify_inode(hammer_inode_t ip, int flags)
1476 {
1477         /* 
1478          * ronly of 0 or 2 does not trigger assertion.
1479          * 2 is a special error state 
1480          */
1481         KKASSERT(ip->hmp->ronly != 1 ||
1482                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1483                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1484                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1485         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1486                 ip->flags |= HAMMER_INODE_RSV_INODES;
1487                 ++ip->hmp->rsv_inodes;
1488         }
1489
1490         ip->flags |= flags;
1491 }
1492
1493 /*
1494  * Request that an inode be flushed.  This whole mess cannot block and may
1495  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1496  * actively flush the inode until the flush can be done.
1497  *
1498  * The inode may already be flushing, or may be in a setup state.  We can
1499  * place the inode in a flushing state if it is currently idle and flag it
1500  * to reflush if it is currently flushing.
1501  *
1502  * Upon return if the inode could not be flushed due to a setup
1503  * dependancy, then it will be automatically flushed when the dependancy
1504  * is satisfied.
1505  */
1506 void
1507 hammer_flush_inode(hammer_inode_t ip, int flags)
1508 {
1509         hammer_mount_t hmp;
1510         hammer_flush_group_t flg;
1511         int good;
1512
1513         /*
1514          * next_flush_group is the first flush group we can place the inode
1515          * in.  It may be NULL.  If it becomes full we append a new flush
1516          * group and make that the next_flush_group.
1517          */
1518         hmp = ip->hmp;
1519         while ((flg = hmp->next_flush_group) != NULL) {
1520                 KKASSERT(flg->running == 0);
1521                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1522                         break;
1523                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1524                 hammer_flusher_async(ip->hmp, flg);
1525         }
1526         if (flg == NULL) {
1527                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1528                 hmp->next_flush_group = flg;
1529                 TAILQ_INIT(&flg->flush_list);
1530                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1531         }
1532
1533         /*
1534          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1535          * state we have to put it back into an IDLE state so we can
1536          * drop the extra ref.
1537          *
1538          * If we have a parent dependancy we must still fall through
1539          * so we can run it.
1540          */
1541         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1542                 if (ip->flush_state == HAMMER_FST_SETUP &&
1543                     TAILQ_EMPTY(&ip->target_list)) {
1544                         ip->flush_state = HAMMER_FST_IDLE;
1545                         hammer_rel_inode(ip, 0);
1546                 }
1547                 if (ip->flush_state == HAMMER_FST_IDLE)
1548                         return;
1549         }
1550
1551         /*
1552          * Our flush action will depend on the current state.
1553          */
1554         switch(ip->flush_state) {
1555         case HAMMER_FST_IDLE:
1556                 /*
1557                  * We have no dependancies and can flush immediately.  Some
1558                  * our children may not be flushable so we have to re-test
1559                  * with that additional knowledge.
1560                  */
1561                 hammer_flush_inode_core(ip, flg, flags);
1562                 break;
1563         case HAMMER_FST_SETUP:
1564                 /*
1565                  * Recurse upwards through dependancies via target_list
1566                  * and start their flusher actions going if possible.
1567                  *
1568                  * 'good' is our connectivity.  -1 means we have none and
1569                  * can't flush, 0 means there weren't any dependancies, and
1570                  * 1 means we have good connectivity.
1571                  */
1572                 good = hammer_setup_parent_inodes(ip, 0, flg);
1573
1574                 if (good >= 0) {
1575                         /*
1576                          * We can continue if good >= 0.  Determine how 
1577                          * many records under our inode can be flushed (and
1578                          * mark them).
1579                          */
1580                         hammer_flush_inode_core(ip, flg, flags);
1581                 } else {
1582                         /*
1583                          * Parent has no connectivity, tell it to flush
1584                          * us as soon as it does.
1585                          *
1586                          * The REFLUSH flag is also needed to trigger
1587                          * dependancy wakeups.
1588                          */
1589                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1590                                      HAMMER_INODE_REFLUSH;
1591                         if (flags & HAMMER_FLUSH_SIGNAL) {
1592                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1593                                 hammer_flusher_async(ip->hmp, flg);
1594                         }
1595                 }
1596                 break;
1597         case HAMMER_FST_FLUSH:
1598                 /*
1599                  * We are already flushing, flag the inode to reflush
1600                  * if needed after it completes its current flush.
1601                  *
1602                  * The REFLUSH flag is also needed to trigger
1603                  * dependancy wakeups.
1604                  */
1605                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1606                         ip->flags |= HAMMER_INODE_REFLUSH;
1607                 if (flags & HAMMER_FLUSH_SIGNAL) {
1608                         ip->flags |= HAMMER_INODE_RESIGNAL;
1609                         hammer_flusher_async(ip->hmp, flg);
1610                 }
1611                 break;
1612         }
1613 }
1614
1615 /*
1616  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1617  * ip which reference our ip.
1618  *
1619  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1620  *     so for now do not ref/deref the structures.  Note that if we use the
1621  *     ref/rel code later, the rel CAN block.
1622  */
1623 static int
1624 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1625                            hammer_flush_group_t flg)
1626 {
1627         hammer_record_t depend;
1628         int good;
1629         int r;
1630
1631         /*
1632          * If we hit our recursion limit and we have parent dependencies
1633          * We cannot continue.  Returning < 0 will cause us to be flagged
1634          * for reflush.  Returning -2 cuts off additional dependency checks
1635          * because they are likely to also hit the depth limit.
1636          *
1637          * We cannot return < 0 if there are no dependencies or there might
1638          * not be anything to wakeup (ip).
1639          */
1640         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1641                 kprintf("HAMMER Warning: depth limit reached on "
1642                         "setup recursion, inode %p %016llx\n",
1643                         ip, (long long)ip->obj_id);
1644                 return(-2);
1645         }
1646
1647         /*
1648          * Scan dependencies
1649          */
1650         good = 0;
1651         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1652                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1653                 KKASSERT(depend->target_ip == ip);
1654                 if (r < 0 && good == 0)
1655                         good = -1;
1656                 if (r > 0)
1657                         good = 1;
1658
1659                 /*
1660                  * If we failed due to the recursion depth limit then stop
1661                  * now.
1662                  */
1663                 if (r == -2)
1664                         break;
1665         }
1666         return(good);
1667 }
1668
1669 /*
1670  * This helper function takes a record representing the dependancy between
1671  * the parent inode and child inode.
1672  *
1673  * record->ip           = parent inode
1674  * record->target_ip    = child inode
1675  * 
1676  * We are asked to recurse upwards and convert the record from SETUP
1677  * to FLUSH if possible.
1678  *
1679  * Return 1 if the record gives us connectivity
1680  *
1681  * Return 0 if the record is not relevant 
1682  *
1683  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1684  */
1685 static int
1686 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1687                                   hammer_flush_group_t flg)
1688 {
1689         hammer_mount_t hmp;
1690         hammer_inode_t pip;
1691         int good;
1692
1693         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1694         pip = record->ip;
1695         hmp = pip->hmp;
1696
1697         /*
1698          * If the record is already flushing, is it in our flush group?
1699          *
1700          * If it is in our flush group but it is a general record or a 
1701          * delete-on-disk, it does not improve our connectivity (return 0),
1702          * and if the target inode is not trying to destroy itself we can't
1703          * allow the operation yet anyway (the second return -1).
1704          */
1705         if (record->flush_state == HAMMER_FST_FLUSH) {
1706                 /*
1707                  * If not in our flush group ask the parent to reflush
1708                  * us as soon as possible.
1709                  */
1710                 if (record->flush_group != flg) {
1711                         pip->flags |= HAMMER_INODE_REFLUSH;
1712                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1713                         return(-1);
1714                 }
1715
1716                 /*
1717                  * If in our flush group everything is already set up,
1718                  * just return whether the record will improve our
1719                  * visibility or not.
1720                  */
1721                 if (record->type == HAMMER_MEM_RECORD_ADD)
1722                         return(1);
1723                 return(0);
1724         }
1725
1726         /*
1727          * It must be a setup record.  Try to resolve the setup dependancies
1728          * by recursing upwards so we can place ip on the flush list.
1729          *
1730          * Limit ourselves to 20 levels of recursion to avoid blowing out
1731          * the kernel stack.  If we hit the recursion limit we can't flush
1732          * until the parent flushes.  The parent will flush independantly
1733          * on its own and ultimately a deep recursion will be resolved.
1734          */
1735         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1736
1737         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1738
1739         /*
1740          * If good < 0 the parent has no connectivity and we cannot safely
1741          * flush the directory entry, which also means we can't flush our
1742          * ip.  Flag us for downward recursion once the parent's
1743          * connectivity is resolved.  Flag the parent for [re]flush or it
1744          * may not check for downward recursions.
1745          */
1746         if (good < 0) {
1747                 pip->flags |= HAMMER_INODE_REFLUSH;
1748                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1749                 return(good);
1750         }
1751
1752         /*
1753          * We are go, place the parent inode in a flushing state so we can
1754          * place its record in a flushing state.  Note that the parent
1755          * may already be flushing.  The record must be in the same flush
1756          * group as the parent.
1757          */
1758         if (pip->flush_state != HAMMER_FST_FLUSH)
1759                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1760         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1761         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1762
1763 #if 0
1764         if (record->type == HAMMER_MEM_RECORD_DEL &&
1765             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1766                 /*
1767                  * Regardless of flushing state we cannot sync this path if the
1768                  * record represents a delete-on-disk but the target inode
1769                  * is not ready to sync its own deletion.
1770                  *
1771                  * XXX need to count effective nlinks to determine whether
1772                  * the flush is ok, otherwise removing a hardlink will
1773                  * just leave the DEL record to rot.
1774                  */
1775                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1776                 return(-1);
1777         } else
1778 #endif
1779         if (pip->flush_group == flg) {
1780                 /*
1781                  * Because we have not calculated nlinks yet we can just
1782                  * set records to the flush state if the parent is in
1783                  * the same flush group as we are.
1784                  */
1785                 record->flush_state = HAMMER_FST_FLUSH;
1786                 record->flush_group = flg;
1787                 ++record->flush_group->refs;
1788                 hammer_ref(&record->lock);
1789
1790                 /*
1791                  * A general directory-add contributes to our visibility.
1792                  *
1793                  * Otherwise it is probably a directory-delete or 
1794                  * delete-on-disk record and does not contribute to our
1795                  * visbility (but we can still flush it).
1796                  */
1797                 if (record->type == HAMMER_MEM_RECORD_ADD)
1798                         return(1);
1799                 return(0);
1800         } else {
1801                 /*
1802                  * If the parent is not in our flush group we cannot
1803                  * flush this record yet, there is no visibility.
1804                  * We tell the parent to reflush and mark ourselves
1805                  * so the parent knows it should flush us too.
1806                  */
1807                 pip->flags |= HAMMER_INODE_REFLUSH;
1808                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1809                 return(-1);
1810         }
1811 }
1812
1813 /*
1814  * This is the core routine placing an inode into the FST_FLUSH state.
1815  */
1816 static void
1817 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1818 {
1819         int go_count;
1820
1821         /*
1822          * Set flush state and prevent the flusher from cycling into
1823          * the next flush group.  Do not place the ip on the list yet.
1824          * Inodes not in the idle state get an extra reference.
1825          */
1826         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1827         if (ip->flush_state == HAMMER_FST_IDLE)
1828                 hammer_ref(&ip->lock);
1829         ip->flush_state = HAMMER_FST_FLUSH;
1830         ip->flush_group = flg;
1831         ++ip->hmp->flusher.group_lock;
1832         ++ip->hmp->count_iqueued;
1833         ++hammer_count_iqueued;
1834         ++flg->total_count;
1835
1836         /*
1837          * If the flush group reaches the autoflush limit we want to signal
1838          * the flusher.  This is particularly important for remove()s.
1839          */
1840         if (flg->total_count == hammer_autoflush)
1841                 flags |= HAMMER_FLUSH_SIGNAL;
1842
1843         /*
1844          * We need to be able to vfsync/truncate from the backend.
1845          */
1846         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1847         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1848                 ip->flags |= HAMMER_INODE_VHELD;
1849                 vref(ip->vp);
1850         }
1851
1852         /*
1853          * Figure out how many in-memory records we can actually flush
1854          * (not including inode meta-data, buffers, etc).
1855          */
1856         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1857         if (flags & HAMMER_FLUSH_RECURSION) {
1858                 /*
1859                  * If this is a upwards recursion we do not want to
1860                  * recurse down again!
1861                  */
1862                 go_count = 1;
1863 #if 0
1864         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1865                 /*
1866                  * No new records are added if we must complete a flush
1867                  * from a previous cycle, but we do have to move the records
1868                  * from the previous cycle to the current one.
1869                  */
1870 #if 0
1871                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1872                                    hammer_syncgrp_child_callback, NULL);
1873 #endif
1874                 go_count = 1;
1875 #endif
1876         } else {
1877                 /*
1878                  * Normal flush, scan records and bring them into the flush.
1879                  * Directory adds and deletes are usually skipped (they are
1880                  * grouped with the related inode rather then with the
1881                  * directory).
1882                  *
1883                  * go_count can be negative, which means the scan aborted
1884                  * due to the flush group being over-full and we should
1885                  * flush what we have.
1886                  */
1887                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1888                                    hammer_setup_child_callback, NULL);
1889         }
1890
1891         /*
1892          * This is a more involved test that includes go_count.  If we
1893          * can't flush, flag the inode and return.  If go_count is 0 we
1894          * were are unable to flush any records in our rec_tree and
1895          * must ignore the XDIRTY flag.
1896          */
1897         if (go_count == 0) {
1898                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1899                         --ip->hmp->count_iqueued;
1900                         --hammer_count_iqueued;
1901
1902                         --flg->total_count;
1903                         ip->flush_state = HAMMER_FST_SETUP;
1904                         ip->flush_group = NULL;
1905                         if (ip->flags & HAMMER_INODE_VHELD) {
1906                                 ip->flags &= ~HAMMER_INODE_VHELD;
1907                                 vrele(ip->vp);
1908                         }
1909
1910                         /*
1911                          * REFLUSH is needed to trigger dependancy wakeups
1912                          * when an inode is in SETUP.
1913                          */
1914                         ip->flags |= HAMMER_INODE_REFLUSH;
1915                         if (flags & HAMMER_FLUSH_SIGNAL) {
1916                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1917                                 hammer_flusher_async(ip->hmp, flg);
1918                         }
1919                         if (--ip->hmp->flusher.group_lock == 0)
1920                                 wakeup(&ip->hmp->flusher.group_lock);
1921                         return;
1922                 }
1923         }
1924
1925         /*
1926          * Snapshot the state of the inode for the backend flusher.
1927          *
1928          * We continue to retain save_trunc_off even when all truncations
1929          * have been resolved as an optimization to determine if we can
1930          * skip the B-Tree lookup for overwrite deletions.
1931          *
1932          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1933          * and stays in ip->flags.  Once set, it stays set until the
1934          * inode is destroyed.
1935          */
1936         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1937                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1938                 ip->sync_trunc_off = ip->trunc_off;
1939                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1940                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1941                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1942
1943                 /*
1944                  * The save_trunc_off used to cache whether the B-Tree
1945                  * holds any records past that point is not used until
1946                  * after the truncation has succeeded, so we can safely
1947                  * set it now.
1948                  */
1949                 if (ip->save_trunc_off > ip->sync_trunc_off)
1950                         ip->save_trunc_off = ip->sync_trunc_off;
1951         }
1952         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1953                            ~HAMMER_INODE_TRUNCATED);
1954         ip->sync_ino_leaf = ip->ino_leaf;
1955         ip->sync_ino_data = ip->ino_data;
1956         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1957 #ifdef DEBUG_TRUNCATE
1958         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1959                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1960 #endif
1961
1962         /*
1963          * The flusher list inherits our inode and reference.
1964          */
1965         KKASSERT(flg->running == 0);
1966         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1967         if (--ip->hmp->flusher.group_lock == 0)
1968                 wakeup(&ip->hmp->flusher.group_lock);
1969
1970         if (flags & HAMMER_FLUSH_SIGNAL) {
1971                 hammer_flusher_async(ip->hmp, flg);
1972         }
1973 }
1974
1975 /*
1976  * Callback for scan of ip->rec_tree.  Try to include each record in our
1977  * flush.  ip->flush_group has been set but the inode has not yet been
1978  * moved into a flushing state.
1979  *
1980  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1981  * both inodes.
1982  *
1983  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1984  * the caller from shortcutting the flush.
1985  */
1986 static int
1987 hammer_setup_child_callback(hammer_record_t rec, void *data)
1988 {
1989         hammer_flush_group_t flg;
1990         hammer_inode_t target_ip;
1991         hammer_inode_t ip;
1992         int r;
1993
1994         /*
1995          * Records deleted or committed by the backend are ignored.
1996          * Note that the flush detects deleted frontend records at
1997          * multiple points to deal with races.  This is just the first
1998          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
1999          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2000          * messes up link-count calculations.
2001          *
2002          * NOTE: Don't get confused between record deletion and, say,
2003          * directory entry deletion.  The deletion of a directory entry
2004          * which is on-media has nothing to do with the record deletion
2005          * flags.
2006          */
2007         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2008                           HAMMER_RECF_COMMITTED)) {
2009                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2010                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2011                         r = 1;
2012                 } else {
2013                         r = 0;
2014                 }
2015                 return(r);
2016         }
2017
2018         /*
2019          * If the record is in an idle state it has no dependancies and
2020          * can be flushed.
2021          */
2022         ip = rec->ip;
2023         flg = ip->flush_group;
2024         r = 0;
2025
2026         switch(rec->flush_state) {
2027         case HAMMER_FST_IDLE:
2028                 /*
2029                  * The record has no setup dependancy, we can flush it.
2030                  */
2031                 KKASSERT(rec->target_ip == NULL);
2032                 rec->flush_state = HAMMER_FST_FLUSH;
2033                 rec->flush_group = flg;
2034                 ++flg->refs;
2035                 hammer_ref(&rec->lock);
2036                 r = 1;
2037                 break;
2038         case HAMMER_FST_SETUP:
2039                 /*
2040                  * The record has a setup dependancy.  These are typically
2041                  * directory entry adds and deletes.  Such entries will be
2042                  * flushed when their inodes are flushed so we do not
2043                  * usually have to add them to the flush here.  However,
2044                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2045                  * it is asking us to flush this record (and it).
2046                  */
2047                 target_ip = rec->target_ip;
2048                 KKASSERT(target_ip != NULL);
2049                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2050
2051                 /*
2052                  * If the target IP is already flushing in our group
2053                  * we could associate the record, but target_ip has
2054                  * already synced ino_data to sync_ino_data and we
2055                  * would also have to adjust nlinks.   Plus there are
2056                  * ordering issues for adds and deletes.
2057                  *
2058                  * Reflush downward if this is an ADD, and upward if
2059                  * this is a DEL.
2060                  */
2061                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2062                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2063                                 ip->flags |= HAMMER_INODE_REFLUSH;
2064                         else
2065                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2066                         break;
2067                 } 
2068
2069                 /*
2070                  * Target IP is not yet flushing.  This can get complex
2071                  * because we have to be careful about the recursion.
2072                  *
2073                  * Directories create an issue for us in that if a flush
2074                  * of a directory is requested the expectation is to flush
2075                  * any pending directory entries, but this will cause the
2076                  * related inodes to recursively flush as well.  We can't
2077                  * really defer the operation so just get as many as we
2078                  * can and
2079                  */
2080 #if 0
2081                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2082                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2083                         /*
2084                          * We aren't reclaiming and the target ip was not
2085                          * previously prevented from flushing due to this
2086                          * record dependancy.  Do not flush this record.
2087                          */
2088                         /*r = 0;*/
2089                 } else
2090 #endif
2091                 if (flg->total_count + flg->refs >
2092                            ip->hmp->undo_rec_limit) {
2093                         /*
2094                          * Our flush group is over-full and we risk blowing
2095                          * out the UNDO FIFO.  Stop the scan, flush what we
2096                          * have, then reflush the directory.
2097                          *
2098                          * The directory may be forced through multiple
2099                          * flush groups before it can be completely
2100                          * flushed.
2101                          */
2102                         ip->flags |= HAMMER_INODE_RESIGNAL |
2103                                      HAMMER_INODE_REFLUSH;
2104                         r = -1;
2105                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2106                         /*
2107                          * If the target IP is not flushing we can force
2108                          * it to flush, even if it is unable to write out
2109                          * any of its own records we have at least one in
2110                          * hand that we CAN deal with.
2111                          */
2112                         rec->flush_state = HAMMER_FST_FLUSH;
2113                         rec->flush_group = flg;
2114                         ++flg->refs;
2115                         hammer_ref(&rec->lock);
2116                         hammer_flush_inode_core(target_ip, flg,
2117                                                 HAMMER_FLUSH_RECURSION);
2118                         r = 1;
2119                 } else {
2120                         /*
2121                          * General or delete-on-disk record.
2122                          *
2123                          * XXX this needs help.  If a delete-on-disk we could
2124                          * disconnect the target.  If the target has its own
2125                          * dependancies they really need to be flushed.
2126                          *
2127                          * XXX
2128                          */
2129                         rec->flush_state = HAMMER_FST_FLUSH;
2130                         rec->flush_group = flg;
2131                         ++flg->refs;
2132                         hammer_ref(&rec->lock);
2133                         hammer_flush_inode_core(target_ip, flg,
2134                                                 HAMMER_FLUSH_RECURSION);
2135                         r = 1;
2136                 }
2137                 break;
2138         case HAMMER_FST_FLUSH:
2139                 /* 
2140                  * The flush_group should already match.
2141                  */
2142                 KKASSERT(rec->flush_group == flg);
2143                 r = 1;
2144                 break;
2145         }
2146         return(r);
2147 }
2148
2149 #if 0
2150 /*
2151  * This version just moves records already in a flush state to the new
2152  * flush group and that is it.
2153  */
2154 static int
2155 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2156 {
2157         hammer_inode_t ip = rec->ip;
2158
2159         switch(rec->flush_state) {
2160         case HAMMER_FST_FLUSH:
2161                 KKASSERT(rec->flush_group == ip->flush_group);
2162                 break;
2163         default:
2164                 break;
2165         }
2166         return(0);
2167 }
2168 #endif
2169
2170 /*
2171  * Wait for a previously queued flush to complete.
2172  *
2173  * If a critical error occured we don't try to wait.
2174  */
2175 void
2176 hammer_wait_inode(hammer_inode_t ip)
2177 {
2178         hammer_flush_group_t flg;
2179
2180         flg = NULL;
2181         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2182                 while (ip->flush_state != HAMMER_FST_IDLE &&
2183                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2184                         if (ip->flush_state == HAMMER_FST_SETUP)
2185                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2186                         if (ip->flush_state != HAMMER_FST_IDLE) {
2187                                 ip->flags |= HAMMER_INODE_FLUSHW;
2188                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2189                         }
2190                 }
2191         }
2192 }
2193
2194 /*
2195  * Called by the backend code when a flush has been completed.
2196  * The inode has already been removed from the flush list.
2197  *
2198  * A pipelined flush can occur, in which case we must re-enter the
2199  * inode on the list and re-copy its fields.
2200  */
2201 void
2202 hammer_flush_inode_done(hammer_inode_t ip, int error)
2203 {
2204         hammer_mount_t hmp;
2205         int dorel;
2206
2207         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2208
2209         hmp = ip->hmp;
2210
2211         /*
2212          * Auto-reflush if the backend could not completely flush
2213          * the inode.  This fixes a case where a deferred buffer flush
2214          * could cause fsync to return early.
2215          */
2216         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2217                 ip->flags |= HAMMER_INODE_REFLUSH;
2218
2219         /*
2220          * Merge left-over flags back into the frontend and fix the state.
2221          * Incomplete truncations are retained by the backend.
2222          */
2223         ip->error = error;
2224         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2225         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2226
2227         /*
2228          * The backend may have adjusted nlinks, so if the adjusted nlinks
2229          * does not match the fronttend set the frontend's RDIRTY flag again.
2230          */
2231         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2232                 ip->flags |= HAMMER_INODE_DDIRTY;
2233
2234         /*
2235          * Fix up the dirty buffer status.
2236          */
2237         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2238                 ip->flags |= HAMMER_INODE_BUFS;
2239         }
2240
2241         /*
2242          * Re-set the XDIRTY flag if some of the inode's in-memory records
2243          * could not be flushed.
2244          */
2245         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2246                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2247                  (!RB_EMPTY(&ip->rec_tree) &&
2248                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2249
2250         /*
2251          * Do not lose track of inodes which no longer have vnode
2252          * assocations, otherwise they may never get flushed again.
2253          *
2254          * The reflush flag can be set superfluously, causing extra pain
2255          * for no reason.  If the inode is no longer modified it no longer
2256          * needs to be flushed.
2257          */
2258         if (ip->flags & HAMMER_INODE_MODMASK) {
2259                 if (ip->vp == NULL)
2260                         ip->flags |= HAMMER_INODE_REFLUSH;
2261         } else {
2262                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2263         }
2264
2265         /*
2266          * Adjust the flush state.
2267          */
2268         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2269                 /*
2270                  * We were unable to flush out all our records, leave the
2271                  * inode in a flush state and in the current flush group.
2272                  * The flush group will be re-run.
2273                  *
2274                  * This occurs if the UNDO block gets too full or there is
2275                  * too much dirty meta-data and allows the flusher to
2276                  * finalize the UNDO block and then re-flush.
2277                  */
2278                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2279                 dorel = 0;
2280         } else {
2281                 /*
2282                  * Remove from the flush_group
2283                  */
2284                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2285                 ip->flush_group = NULL;
2286
2287                 /*
2288                  * Clean up the vnode ref and tracking counts.
2289                  */
2290                 if (ip->flags & HAMMER_INODE_VHELD) {
2291                         ip->flags &= ~HAMMER_INODE_VHELD;
2292                         vrele(ip->vp);
2293                 }
2294                 --hmp->count_iqueued;
2295                 --hammer_count_iqueued;
2296
2297                 /*
2298                  * And adjust the state.
2299                  */
2300                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2301                         ip->flush_state = HAMMER_FST_IDLE;
2302                         dorel = 1;
2303                 } else {
2304                         ip->flush_state = HAMMER_FST_SETUP;
2305                         dorel = 0;
2306                 }
2307
2308                 /*
2309                  * If the frontend is waiting for a flush to complete,
2310                  * wake it up.
2311                  */
2312                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2313                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2314                         wakeup(&ip->flags);
2315                 }
2316
2317                 /*
2318                  * If the frontend made more changes and requested another
2319                  * flush, then try to get it running.
2320                  *
2321                  * Reflushes are aborted when the inode is errored out.
2322                  */
2323                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2324                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2325                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2326                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2327                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2328                         } else {
2329                                 hammer_flush_inode(ip, 0);
2330                         }
2331                 }
2332         }
2333
2334         /*
2335          * If we have no parent dependancies we can clear CONN_DOWN
2336          */
2337         if (TAILQ_EMPTY(&ip->target_list))
2338                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2339
2340         /*
2341          * If the inode is now clean drop the space reservation.
2342          */
2343         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2344             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2345                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2346                 --hmp->rsv_inodes;
2347         }
2348
2349         if (dorel)
2350                 hammer_rel_inode(ip, 0);
2351 }
2352
2353 /*
2354  * Called from hammer_sync_inode() to synchronize in-memory records
2355  * to the media.
2356  */
2357 static int
2358 hammer_sync_record_callback(hammer_record_t record, void *data)
2359 {
2360         hammer_cursor_t cursor = data;
2361         hammer_transaction_t trans = cursor->trans;
2362         hammer_mount_t hmp = trans->hmp;
2363         int error;
2364
2365         /*
2366          * Skip records that do not belong to the current flush.
2367          */
2368         ++hammer_stats_record_iterations;
2369         if (record->flush_state != HAMMER_FST_FLUSH)
2370                 return(0);
2371
2372 #if 1
2373         if (record->flush_group != record->ip->flush_group) {
2374                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2375                 Debugger("blah2");
2376                 return(0);
2377         }
2378 #endif
2379         KKASSERT(record->flush_group == record->ip->flush_group);
2380
2381         /*
2382          * Interlock the record using the BE flag.  Once BE is set the
2383          * frontend cannot change the state of FE.
2384          *
2385          * NOTE: If FE is set prior to us setting BE we still sync the
2386          * record out, but the flush completion code converts it to 
2387          * a delete-on-disk record instead of destroying it.
2388          */
2389         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2390         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2391
2392         /*
2393          * The backend has already disposed of the record.
2394          */
2395         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2396                 error = 0;
2397                 goto done;
2398         }
2399
2400         /*
2401          * If the whole inode is being deleting all on-disk records will
2402          * be deleted very soon, we can't sync any new records to disk
2403          * because they will be deleted in the same transaction they were
2404          * created in (delete_tid == create_tid), which will assert.
2405          *
2406          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2407          * that we currently panic on.
2408          */
2409         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2410                 switch(record->type) {
2411                 case HAMMER_MEM_RECORD_DATA:
2412                         /*
2413                          * We don't have to do anything, if the record was
2414                          * committed the space will have been accounted for
2415                          * in the blockmap.
2416                          */
2417                         /* fall through */
2418                 case HAMMER_MEM_RECORD_GENERAL:
2419                         /*
2420                          * Set deleted-by-backend flag.  Do not set the
2421                          * backend committed flag, because we are throwing
2422                          * the record away.
2423                          */
2424                         record->flags |= HAMMER_RECF_DELETED_BE;
2425                         ++record->ip->rec_generation;
2426                         error = 0;
2427                         goto done;
2428                 case HAMMER_MEM_RECORD_ADD:
2429                         panic("hammer_sync_record_callback: illegal add "
2430                               "during inode deletion record %p", record);
2431                         break; /* NOT REACHED */
2432                 case HAMMER_MEM_RECORD_INODE:
2433                         panic("hammer_sync_record_callback: attempt to "
2434                               "sync inode record %p?", record);
2435                         break; /* NOT REACHED */
2436                 case HAMMER_MEM_RECORD_DEL:
2437                         /* 
2438                          * Follow through and issue the on-disk deletion
2439                          */
2440                         break;
2441                 }
2442         }
2443
2444         /*
2445          * If DELETED_FE is set special handling is needed for directory
2446          * entries.  Dependant pieces related to the directory entry may
2447          * have already been synced to disk.  If this occurs we have to
2448          * sync the directory entry and then change the in-memory record
2449          * from an ADD to a DELETE to cover the fact that it's been
2450          * deleted by the frontend.
2451          *
2452          * A directory delete covering record (MEM_RECORD_DEL) can never
2453          * be deleted by the frontend.
2454          *
2455          * Any other record type (aka DATA) can be deleted by the frontend.
2456          * XXX At the moment the flusher must skip it because there may
2457          * be another data record in the flush group for the same block,
2458          * meaning that some frontend data changes can leak into the backend's
2459          * synchronization point.
2460          */
2461         if (record->flags & HAMMER_RECF_DELETED_FE) {
2462                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2463                         /*
2464                          * Convert a front-end deleted directory-add to
2465                          * a directory-delete entry later.
2466                          */
2467                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2468                 } else {
2469                         /*
2470                          * Dispose of the record (race case).  Mark as
2471                          * deleted by backend (and not committed).
2472                          */
2473                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2474                         record->flags |= HAMMER_RECF_DELETED_BE;
2475                         ++record->ip->rec_generation;
2476                         error = 0;
2477                         goto done;
2478                 }
2479         }
2480
2481         /*
2482          * Assign the create_tid for new records.  Deletions already
2483          * have the record's entire key properly set up.
2484          */
2485         if (record->type != HAMMER_MEM_RECORD_DEL) {
2486                 record->leaf.base.create_tid = trans->tid;
2487                 record->leaf.create_ts = trans->time32;
2488         }
2489         for (;;) {
2490                 error = hammer_ip_sync_record_cursor(cursor, record);
2491                 if (error != EDEADLK)
2492                         break;
2493                 hammer_done_cursor(cursor);
2494                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2495                                            record->ip);
2496                 if (error)
2497                         break;
2498         }
2499         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2500
2501         if (error)
2502                 error = -error;
2503 done:
2504         hammer_flush_record_done(record, error);
2505
2506         /*
2507          * Do partial finalization if we have built up too many dirty
2508          * buffers.  Otherwise a buffer cache deadlock can occur when
2509          * doing things like creating tens of thousands of tiny files.
2510          *
2511          * We must release our cursor lock to avoid a 3-way deadlock
2512          * due to the exclusive sync lock the finalizer must get.
2513          */
2514         if (hammer_flusher_meta_limit(hmp)) {
2515                 hammer_unlock_cursor(cursor);
2516                 hammer_flusher_finalize(trans, 0);
2517                 hammer_lock_cursor(cursor);
2518         }
2519
2520         return(error);
2521 }
2522
2523 /*
2524  * Backend function called by the flusher to sync an inode to media.
2525  */
2526 int
2527 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2528 {
2529         struct hammer_cursor cursor;
2530         hammer_node_t tmp_node;
2531         hammer_record_t depend;
2532         hammer_record_t next;
2533         int error, tmp_error;
2534         u_int64_t nlinks;
2535
2536         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2537                 return(0);
2538
2539         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2540         if (error)
2541                 goto done;
2542
2543         /*
2544          * Any directory records referencing this inode which are not in
2545          * our current flush group must adjust our nlink count for the
2546          * purposes of synchronization to disk.
2547          *
2548          * Records which are in our flush group can be unlinked from our
2549          * inode now, potentially allowing the inode to be physically
2550          * deleted.
2551          *
2552          * This cannot block.
2553          */
2554         nlinks = ip->ino_data.nlinks;
2555         next = TAILQ_FIRST(&ip->target_list);
2556         while ((depend = next) != NULL) {
2557                 next = TAILQ_NEXT(depend, target_entry);
2558                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2559                     depend->flush_group == ip->flush_group) {
2560                         /*
2561                          * If this is an ADD that was deleted by the frontend
2562                          * the frontend nlinks count will have already been
2563                          * decremented, but the backend is going to sync its
2564                          * directory entry and must account for it.  The
2565                          * record will be converted to a delete-on-disk when
2566                          * it gets synced.
2567                          *
2568                          * If the ADD was not deleted by the frontend we
2569                          * can remove the dependancy from our target_list.
2570                          */
2571                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2572                                 ++nlinks;
2573                         } else {
2574                                 TAILQ_REMOVE(&ip->target_list, depend,
2575                                              target_entry);
2576                                 depend->target_ip = NULL;
2577                         }
2578                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2579                         /*
2580                          * Not part of our flush group and not deleted by
2581                          * the front-end, adjust the link count synced to
2582                          * the media (undo what the frontend did when it
2583                          * queued the record).
2584                          */
2585                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2586                         switch(depend->type) {
2587                         case HAMMER_MEM_RECORD_ADD:
2588                                 --nlinks;
2589                                 break;
2590                         case HAMMER_MEM_RECORD_DEL:
2591                                 ++nlinks;
2592                                 break;
2593                         default:
2594                                 break;
2595                         }
2596                 }
2597         }
2598
2599         /*
2600          * Set dirty if we had to modify the link count.
2601          */
2602         if (ip->sync_ino_data.nlinks != nlinks) {
2603                 KKASSERT((int64_t)nlinks >= 0);
2604                 ip->sync_ino_data.nlinks = nlinks;
2605                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2606         }
2607
2608         /*
2609          * If there is a trunction queued destroy any data past the (aligned)
2610          * truncation point.  Userland will have dealt with the buffer
2611          * containing the truncation point for us.
2612          *
2613          * We don't flush pending frontend data buffers until after we've
2614          * dealt with the truncation.
2615          */
2616         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2617                 /*
2618                  * Interlock trunc_off.  The VOP front-end may continue to
2619                  * make adjustments to it while we are blocked.
2620                  */
2621                 off_t trunc_off;
2622                 off_t aligned_trunc_off;
2623                 int blkmask;
2624
2625                 trunc_off = ip->sync_trunc_off;
2626                 blkmask = hammer_blocksize(trunc_off) - 1;
2627                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2628
2629                 /*
2630                  * Delete any whole blocks on-media.  The front-end has
2631                  * already cleaned out any partial block and made it
2632                  * pending.  The front-end may have updated trunc_off
2633                  * while we were blocked so we only use sync_trunc_off.
2634                  *
2635                  * This operation can blow out the buffer cache, EWOULDBLOCK
2636                  * means we were unable to complete the deletion.  The
2637                  * deletion will update sync_trunc_off in that case.
2638                  */
2639                 error = hammer_ip_delete_range(&cursor, ip,
2640                                                 aligned_trunc_off,
2641                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2642                 if (error == EWOULDBLOCK) {
2643                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2644                         error = 0;
2645                         goto defer_buffer_flush;
2646                 }
2647
2648                 if (error)
2649                         goto done;
2650
2651                 /*
2652                  * Clear the truncation flag on the backend after we have
2653                  * complete the deletions.  Backend data is now good again
2654                  * (including new records we are about to sync, below).
2655                  *
2656                  * Leave sync_trunc_off intact.  As we write additional
2657                  * records the backend will update sync_trunc_off.  This
2658                  * tells the backend whether it can skip the overwrite
2659                  * test.  This should work properly even when the backend
2660                  * writes full blocks where the truncation point straddles
2661                  * the block because the comparison is against the base
2662                  * offset of the record.
2663                  */
2664                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2665                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2666         } else {
2667                 error = 0;
2668         }
2669
2670         /*
2671          * Now sync related records.  These will typically be directory
2672          * entries, records tracking direct-writes, or delete-on-disk records.
2673          */
2674         if (error == 0) {
2675                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2676                                     hammer_sync_record_callback, &cursor);
2677                 if (tmp_error < 0)
2678                         tmp_error = -error;
2679                 if (tmp_error)
2680                         error = tmp_error;
2681         }
2682         hammer_cache_node(&ip->cache[1], cursor.node);
2683
2684         /*
2685          * Re-seek for inode update, assuming our cache hasn't been ripped
2686          * out from under us.
2687          */
2688         if (error == 0) {
2689                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2690                 if (tmp_node) {
2691                         hammer_cursor_downgrade(&cursor);
2692                         hammer_lock_sh(&tmp_node->lock);
2693                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2694                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2695                         hammer_unlock(&tmp_node->lock);
2696                         hammer_rel_node(tmp_node);
2697                 }
2698                 error = 0;
2699         }
2700
2701         /*
2702          * If we are deleting the inode the frontend had better not have
2703          * any active references on elements making up the inode.
2704          *
2705          * The call to hammer_ip_delete_clean() cleans up auxillary records
2706          * but not DB or DATA records.  Those must have already been deleted
2707          * by the normal truncation mechanic.
2708          */
2709         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2710                 RB_EMPTY(&ip->rec_tree)  &&
2711             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2712             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2713                 int count1 = 0;
2714
2715                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2716                 if (error == 0) {
2717                         ip->flags |= HAMMER_INODE_DELETED;
2718                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2719                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2720                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2721
2722                         /*
2723                          * Set delete_tid in both the frontend and backend
2724                          * copy of the inode record.  The DELETED flag handles
2725                          * this, do not set RDIRTY.
2726                          */
2727                         ip->ino_leaf.base.delete_tid = trans->tid;
2728                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2729                         ip->ino_leaf.delete_ts = trans->time32;
2730                         ip->sync_ino_leaf.delete_ts = trans->time32;
2731
2732
2733                         /*
2734                          * Adjust the inode count in the volume header
2735                          */
2736                         hammer_sync_lock_sh(trans);
2737                         if (ip->flags & HAMMER_INODE_ONDISK) {
2738                                 hammer_modify_volume_field(trans,
2739                                                            trans->rootvol,
2740                                                            vol0_stat_inodes);
2741                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2742                                 hammer_modify_volume_done(trans->rootvol);
2743                         }
2744                         hammer_sync_unlock(trans);
2745                 }
2746         }
2747
2748         if (error)
2749                 goto done;
2750         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2751
2752 defer_buffer_flush:
2753         /*
2754          * Now update the inode's on-disk inode-data and/or on-disk record.
2755          * DELETED and ONDISK are managed only in ip->flags.
2756          *
2757          * In the case of a defered buffer flush we still update the on-disk
2758          * inode to satisfy visibility requirements if there happen to be
2759          * directory dependancies.
2760          */
2761         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2762         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2763                 /*
2764                  * If deleted and on-disk, don't set any additional flags.
2765                  * the delete flag takes care of things.
2766                  *
2767                  * Clear flags which may have been set by the frontend.
2768                  */
2769                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2770                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2771                                     HAMMER_INODE_DELETING);
2772                 break;
2773         case HAMMER_INODE_DELETED:
2774                 /*
2775                  * Take care of the case where a deleted inode was never
2776                  * flushed to the disk in the first place.
2777                  *
2778                  * Clear flags which may have been set by the frontend.
2779                  */
2780                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2781                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2782                                     HAMMER_INODE_DELETING);
2783                 while (RB_ROOT(&ip->rec_tree)) {
2784                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2785                         hammer_ref(&record->lock);
2786                         KKASSERT(record->lock.refs == 1);
2787                         record->flags |= HAMMER_RECF_DELETED_BE;
2788                         ++record->ip->rec_generation;
2789                         hammer_rel_mem_record(record);
2790                 }
2791                 break;
2792         case HAMMER_INODE_ONDISK:
2793                 /*
2794                  * If already on-disk, do not set any additional flags.
2795                  */
2796                 break;
2797         default:
2798                 /*
2799                  * If not on-disk and not deleted, set DDIRTY to force
2800                  * an initial record to be written.
2801                  *
2802                  * Also set the create_tid in both the frontend and backend
2803                  * copy of the inode record.
2804                  */
2805                 ip->ino_leaf.base.create_tid = trans->tid;
2806                 ip->ino_leaf.create_ts = trans->time32;
2807                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2808                 ip->sync_ino_leaf.create_ts = trans->time32;
2809                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2810                 break;
2811         }
2812
2813         /*
2814          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2815          * is already on-disk the old record is marked as deleted.
2816          *
2817          * If DELETED is set hammer_update_inode() will delete the existing
2818          * record without writing out a new one.
2819          *
2820          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2821          */
2822         if (ip->flags & HAMMER_INODE_DELETED) {
2823                 error = hammer_update_inode(&cursor, ip);
2824         } else 
2825         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2826             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2827                 error = hammer_update_itimes(&cursor, ip);
2828         } else
2829         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2830                 error = hammer_update_inode(&cursor, ip);
2831         }
2832 done:
2833         if (error) {
2834                 hammer_critical_error(ip->hmp, ip, error,
2835                                       "while syncing inode");
2836         }
2837         hammer_done_cursor(&cursor);
2838         return(error);
2839 }
2840
2841 /*
2842  * This routine is called when the OS is no longer actively referencing
2843  * the inode (but might still be keeping it cached), or when releasing
2844  * the last reference to an inode.
2845  *
2846  * At this point if the inode's nlinks count is zero we want to destroy
2847  * it, which may mean destroying it on-media too.
2848  */
2849 void
2850 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2851 {
2852         struct vnode *vp;
2853
2854         /*
2855          * Set the DELETING flag when the link count drops to 0 and the
2856          * OS no longer has any opens on the inode.
2857          *
2858          * The backend will clear DELETING (a mod flag) and set DELETED
2859          * (a state flag) when it is actually able to perform the
2860          * operation.
2861          *
2862          * Don't reflag the deletion if the flusher is currently syncing
2863          * one that was already flagged.  A previously set DELETING flag
2864          * may bounce around flags and sync_flags until the operation is
2865          * completely done.
2866          */
2867         if (ip->ino_data.nlinks == 0 &&
2868             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2869                 ip->flags |= HAMMER_INODE_DELETING;
2870                 ip->flags |= HAMMER_INODE_TRUNCATED;
2871                 ip->trunc_off = 0;
2872                 vp = NULL;
2873                 if (getvp) {
2874                         if (hammer_get_vnode(ip, &vp) != 0)
2875                                 return;
2876                 }
2877
2878                 /*
2879                  * Final cleanup
2880                  */
2881                 if (ip->vp) {
2882                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2883                         vnode_pager_setsize(ip->vp, 0);
2884                 }
2885                 if (getvp) {
2886                         vput(vp);
2887                 }
2888         }
2889 }
2890
2891 /*
2892  * After potentially resolving a dependancy the inode is tested
2893  * to determine whether it needs to be reflushed.
2894  */
2895 void
2896 hammer_test_inode(hammer_inode_t ip)
2897 {
2898         if (ip->flags & HAMMER_INODE_REFLUSH) {
2899                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2900                 hammer_ref(&ip->lock);
2901                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2902                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2903                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2904                 } else {
2905                         hammer_flush_inode(ip, 0);
2906                 }
2907                 hammer_rel_inode(ip, 0);
2908         }
2909 }
2910
2911 /*
2912  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2913  * reassociated with a vp or just before it gets freed.
2914  *
2915  * Pipeline wakeups to threads blocked due to an excessive number of
2916  * detached inodes.  The reclaim count generates a bit of negative
2917  * feedback.
2918  */
2919 static void
2920 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2921 {
2922         struct hammer_reclaim *reclaim;
2923         hammer_mount_t hmp = ip->hmp;
2924
2925         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2926                 return;
2927
2928         --hammer_count_reclaiming;
2929         --hmp->inode_reclaims;
2930         ip->flags &= ~HAMMER_INODE_RECLAIM;
2931
2932         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2933                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
2934                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
2935                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2936                         wakeup(reclaim);
2937                 }
2938         }
2939 }
2940
2941 /*
2942  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2943  * inodes build up before we start blocking.
2944  *
2945  * When we block we don't care *which* inode has finished reclaiming,
2946  * as lone as one does.  This is somewhat heuristical... we also put a
2947  * cap on how long we are willing to wait.
2948  */
2949 void
2950 hammer_inode_waitreclaims(hammer_mount_t hmp)
2951 {
2952         struct hammer_reclaim reclaim;
2953         int delay;
2954
2955         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
2956                 return;
2957         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2958                 (HAMMER_RECLAIM_WAIT * 3) + 1;
2959         if (delay > 0) {
2960                 reclaim.count = 2;
2961                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
2962                 tsleep(&reclaim, 0, "hmrrcm", delay);
2963                 if (reclaim.count > 0)
2964                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2965         }
2966 }
2967
2968 /*
2969  * A larger then normal backlog of inodes is sitting in the flusher,
2970  * enforce a general slowdown to let it catch up.  This routine is only
2971  * called on completion of a non-flusher-related transaction which
2972  * performed B-Tree node I/O.
2973  *
2974  * It is possible for the flusher to stall in a continuous load.
2975  * blogbench -i1000 -o seems to do a good job generating this sort of load.
2976  * If the flusher is unable to catch up the inode count can bloat until
2977  * we run out of kvm.
2978  *
2979  * This is a bit of a hack.
2980  */
2981 void
2982 hammer_inode_waithard(hammer_mount_t hmp)
2983 {
2984         /*
2985          * Hysteresis.
2986          */
2987         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
2988                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
2989                     hmp->count_iqueued < hmp->count_inodes / 20) {
2990                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
2991                         return;
2992                 }
2993         } else {
2994                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
2995                     hmp->count_iqueued < hmp->count_inodes / 10) {
2996                         return;
2997                 }
2998                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
2999         }
3000
3001         /*
3002          * Block for one flush cycle.
3003          */
3004         hammer_flusher_wait_next(hmp);
3005 }
3006